#1077 feat(arm): build and run Strata on aarch64, including DGX Spark

Beschreibung

## Summary
- Builds and runs on aarch64. On a DGX Spark (GB10, sm_121) the engine is compiled in Docker, every expert stays in the unified-memory cache, and setup does not add a RAM budget.
- Integrated-GPU cache sizing was already in 0.1.40. This is the aarch64 build, the NEON CPU path, and the Spark setup/Docker path.
- x86 builds are unchanged. AVX kernels are not compiled on ARM.

## Test plan
- [x] x86 setup unit tests (golden, Unsloth, older GPUs, Strix Halo, AMD)
- [x] DGX Spark, Docker image `strata`: `uname -m` is aarch64, binary is a 54 MB AArch64 ELF, `BUILD.json` archs `[121]`, vision none
- [x] `UD-IQ4_XS`, context 262144: GPU line `NVIDIA GB10, 122 GB unified memory, compute capability 12.1`; `unified memory: no RAM budget`; expert cache `24576 experts, 55.43 GiB`; server ready on `:8080`
- [x] Chat reply: 37 tokens at 47.0 tok/s
- [ ] BetterBench still running; numbers to follow